Grazing plays a key role in reducing the external inputs required for ruminant production and in alleviating feed-food competition. Beyond the production of meat and milk, grassland-based systems provide a wide range of ecosystem services. Agroecology and organic farming aim to reconcile natural resource management and food production, in the long term, based on the management of ecological processes. In this perspective paper, we report what we have learned from case studies with beef cattle, sheep, and dairy cattle across Uruguay and western Europe, in which we have been involved. Multicriteria methods, such as Pareto frontiers and positive deviances, were used to analyze trade-offs and identify win-wins from farm surveys. Long-term farm networks coupled with bioeconomic optimization models revealed fluctuations in farm income and allowed estimating system resilience. Extensive farmlet experiments made it possible to integrate knowledge on animal physiology and grassland ecology in the system redesign process and to test for innovative and risky management options that could lead to unacceptable learning costs in commercial farms. Finally, learning from farmers' local knowledge in teams with researchers and technical advisers can provide positive changes in grazing systems. In Uruguayan family farms, for example, the scientific knowledge gained from farmlet experiments led to advice on management options based on farm-specific diagnosis. Farmers adapted the proposals, with researchers supporting the processes by providing quantitative information on consequences and spaces for reflection. In a French cheese production area, the focus was on farmers' own experience. Games facilitated interactions as participants could challenge each other's reasoning and conclusions in a safe environment. These two case studies illustrate the diversity of co-innovation approaches, but in both cases knowledge sharing between researchers, farmers, and other stakeholders appeared more efficient to help farmers understand and adapt their own system properties than researching "best practice" solutions for large-scale transfer.